The Unmixed Reality Knowledge Injection via Merge: Efficient Domain Adaptation Without Post-Training

The dominant approach for creating domain-specialized large language models requires expensive post-training procedures including supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and instruction tuning. The best performing models undergo multiple stages of alignment after pretraining on domain-specific corpora, with costs often exceeding initial pretraining. We propose a new efficient paradigm, Knowledge Injection via Merge (KIM), that dispenses with post-training on the domain model entirely by leveraging asymmetric model merging guided by mechanistic interpretability insights. Given a domain corpus, we pretrain a knowledge model and merge it with an already-aligned general model using 𝑀 𝐵 𝑀 𝐴 layer-specific interpolation that preserves alignment-critical parameters while injecting domain knowledge into factual storage layers. Experiments on medical, legal, and scientific domains show our approach achieves 94.2% of the knowledge retention of a fully fine-tuned model on domain-specific benchmarks while preserving 97.8% of the alignment capabilities, using less than 15% of the total compute budget. We show that KIM generalizes across model scales (7B-70B parameters) and architectures (Llama, Mistral), enabling efficient domain adaptation for practitioners without access to expensive alignment infrastructure.